MCP server for deploying apps: what to look for
An MCP deploy server lets Claude Code, Cursor, Codex or Copilot deploy, read logs and roll back on your hosting account. Judge one by how it signs in, how far a key reaches, whether its tools say which ones change things, and what you can see and revoke afterwards. Vercel, Render, Railway, Netlify and Livops all ship one.
By Livops team, the people who build and run LivopsUpdated
Most hosting platforms now let an AI coding agent run deploys through the Model Context Protocol. They are not the same thing with different logos: they differ in how the agent signs in, how much a leaked key can do, whether the client is told which tools change things, and what you can see and switch off afterwards. This page lists what to check, shows how Vercel, Render, Railway and Netlify answer from their own docs, and explains how the Livops server works.
What is an MCP deploy server?
MCP is an open standard for letting a language model call another service’s tools. The model discovers the tools, picks one, and the client sends the call[1]. A deploy server’s tools are a hosting dashboard’s buttons: list projects, deploy, read the build log, set a variable, roll back.
The server is only as safe as the checks behind each tool. The protocol itself asks two things of the client: there should always be a human able to deny a tool call, and annotations such as destructiveHint must be treated as untrusted unless they come from a trusted server[1].
What should you look for?
How it signs in
The MCP authorization spec is OAuth based[2]: the client is sent to the provider’s own sign-in page, you approve it, and the client gets a short-lived key it never had to copy. Prefer that whenever a person is at the keyboard. An API key pasted into a config file is the fallback for CI and for clients without OAuth; check where the client stores it and whether the provider lets you make a key that is narrower than your account.
How far one key reaches
Ask what the key is scoped to: one team or every team, your role or more, and whether a key meant for CI can do what a signed-in person can. A key that sits in a CI secret is the one most likely to leak, and it should be the one that can do least.
Whether the tools say what they do
Each MCP tool may carry readOnlyHint and destructiveHint. A client uses them to decide when to ask you. A server where every tool that changes something says so lets you approve writes and let reads through without prompts.
What you can see and revoke
Look for a list of connected apps with a disconnect button, a record of who deployed or changed what, and rate limits. The MCP specification asks servers to rate limit tool calls and clients to log them[1].
What the model reads
Logs, commit messages and database rows go to the model as text. A log line that says “ignore previous instructions” is data, but the model still reads it. The client’s approval prompt is what stops a bad instruction from becoming an action, so keep it on for writes.
The options at a glance
| Server | Address | Sign-in | Deletes through MCP |
|---|---|---|---|
| Livops | https://livops.uz/api/mcp |
Browser OAuth, or an lvp_ API token |
Browser sign-in, as your team role allows; never with a token |
| Vercel | https://mcp.vercel.com[3] |
OAuth; clients must use CIMD or be reviewed by Vercel[3] | Not stated on the setup page; access equals your Vercel account[3] |
| Render | https://mcp.render.com/mcp[5] |
OAuth or an API key[5] | Not offered; use the Dashboard or REST API[5] |
| Railway | https://mcp.railway.com, or railway mcp from the CLI[6] |
OAuth, or your CLI login; project tokens refused[6] | Some tools, marked destructive[6] |
| Netlify | https://mcp.netlify.com/mcp, or local @netlify/mcp[7] |
Sign in to Netlify when the agent gives you a link[7] | The setup page does not list tools[7] |
Facts about the other servers are from each provider’s own documentation, read on 19 November 2026.
Each option
Vercel
A remote server with OAuth that implements the MCP authorization and Streamable HTTP specs[3]. It searches the docs, manages projects and deployments, analyses deployment logs and queries Web Analytics[3], and its tool reference covers far more categories than that, from environment variables and domains to firewall and storage[4]. Connecting it gives the AI the same access as your Vercel user account, and Vercel’s own advice is to enable human confirmation[3]. It is stricter than most about which clients may connect: a client must use a Client ID Metadata Document or be reviewed and approved[3]. Its list of documented clients is also longer than anyone else’s here.
Render
A hosted server with OAuth or an API key; with a key, the AI reaches only the workspace and permissions you granted that key[5]. It can create web services, static sites, cron jobs, Postgres and Key Value, trigger deploys, update a service’s environment variables, read logs and metrics, and run read-only SQL[5]. Render is clear about its edge: other modifications and deletions go through the Dashboard or REST API[5], and it warns that what remains includes potentially destructive operations such as changing environment variables[5]. A server that cannot delete is a clean answer to the hardest question on this page.
Railway
A hosted server, reached through railway mcp in the CLI or directly over OAuth, where you choose which workspaces and projects the client may use; tokens are short-lived and revocable[6]. Project tokens are not accepted, because the server needs a person for billing and audit trails[6]. Destructive tools are marked at the protocol level so clients ask first, and the local server returns a preview and needs confirm: true before acting[6]. Choosing projects on the consent screen is finer than Livops, where a connection is one whole team.
Netlify
Netlify recommends its remote server and keeps a local one for environments that require it[7]. You sign in to Netlify when the agent hands you a link[7]. The setup page we read lists clients and install commands, not tools, so we do not describe its tools here.
How the Livops MCP server works
One address, https://livops.uz/api/mcp, over Streamable HTTP: one POST, one JSON answer, no session and no server-sent stream. It speaks protocol versions 2025-11-25, 2025-06-18 and 2025-03-26. The CLI serves the same tools over stdio with livops mcp.
Sign-in. Browser sign-in is an OAuth 2.1 server for this endpoint only. Clients register themselves, PKCE is required with S256, and the key an AI client gets is issued for /api/mcp and honoured nowhere else. On the consent screen you see the app’s name, pick one team, see what it will be able to do, and press Allow; with two-factor on, your code comes first. The access key lasts an hour and renews itself. The app’s name is the app’s own and is not verified, so a request you did not start gets Cancel.
Reach. A connection is one team and one person, with that person’s role. A viewer only reads. The connection can move to another team you belong to with switch_team, and then has your role there.
Two kinds of key. Through the browser, the client can do what your role allows: create, change, stop, delete, read a database password or a bucket key, redeem a promo code, open a Payme or Click top-up link. An lvp_ API token can deploy and read, and the server refuses it anything that deletes, reveals a password or a bucket key, pays, or rewrites the variables a whole environment shares, whichever tool sends the request. Some things have no tool at all: card numbers, your password, two-factor, sessions, creating API tokens, deleting the account, handing a project over, project members, downloading a raw backup.
Hints. Every read-only tool sets readOnlyHint; every tool that changes something sets destructiveHint, and a server test fails if one does not. A client that asks before writes therefore asks before every Livops write.
Every call is an ordinary API call. A tool never reaches a service directly. It is sent back through the API’s own router with your key, so it passes the same token, team, role and project-owner checks as a request from the CLI or the cabinet. Calls are rate limited per account.
What you can see and undo. Connected clients are listed under Account → Connected AI apps, with Disconnect. Tokens are listed under Settings → API tokens with when each was last used. The team’s Activity log shows who deployed, changed variables or changed members, and marks what came through a token or a connected app as “via token”.
What it does not hand over. list_variables returns names and which are secret, never values. list_databases returns no passwords. query_database is the team owner’s only, and a statement that writes needs allow_write.
Which should you pick?
Pick by where your app runs, not by the MCP server: every server here deploys to its own platform only. If you are already on Vercel, Render, Railway or Netlify, use theirs. If your users are in Uzbekistan and you pay in so’m, Livops is the one whose servers are there. Either way, connect through the browser when you are at the keyboard, give CI its own narrow key, and leave the client’s confirmation on for anything that changes.
The step-by-step setup for each client is in Deploy from Claude Code, Cursor, Codex CLI and VS Code. Every Livops tool is listed in MCP (AI clients).
Questions
What is an MCP server for deploys?
A service that offers your hosting account to an AI client as tools: list projects, deploy, read the build log, roll back. The client's model picks the tool and the server runs it as you, through the same permission checks as the dashboard.
OAuth or an API key?
OAuth in the browser when a person is at the keyboard: nothing to copy, short-lived keys, and one place to disconnect. A key for CI and for clients that cannot sign in. On Livops the key is also narrower: an lvp_ token cannot delete, read passwords or pay.
Do destructive hints make a server safe?
No. readOnlyHint and destructiveHint tell the client which calls to ask about, and the MCP specification says clients must treat them as untrusted unless the server is trusted. The safety comes from the server's own permission checks and from you approving the calls that change things.
Can an AI client delete my production app through MCP?
It depends on the server and the key. Render's MCP server does not offer deletions. Railway marks its destructive tools so clients ask first. On Livops, a browser connection can delete what your team role allows, and the client asks before it runs that tool; an lvp_ token cannot delete anything.
Is Livops's MCP server listed in Claude's Connectors Directory?
No. Add it yourself as a custom connector with the address https://livops.uz/api/mcp, or add that address to Claude Code, Cursor, Codex, VS Code, Gemini CLI or Zed.
Where are Livops's servers?
In Uzbekistan. Credits are bought in so'm through Payme or Click, with no card required. If your users are in the US or Europe, a platform with regions there will be closer to them.
Sources
- Model Context Protocol — Tools (specification 2025-11-25) (modelcontextprotocol.io). Checked .
- Model Context Protocol — Authorization (specification 2025-11-25) (modelcontextprotocol.io). Checked .
- Vercel — Use Vercel MCP (vercel.com). Checked .
- Vercel — Vercel MCP tools (vercel.com). Checked .
- Render — Render MCP Server (render.com). Checked .
- Railway — Railway MCP Server (docs.railway.com). Checked .
- Netlify — Agent setup overview (docs.netlify.com). Checked .
Put your project online today
Connect GitHub or drop a folder. No card is asked for.
Connect your AI clientRelated
- Deploy from Claude Code: connect Livops over MCPRun one command, claude mcp add --transport http livops https://livops.uz/api/mcp, then sign in through /mcp. From then on Claude Code can list your projects, deploy, read the build log, explain a failure and roll back, as you and only in the team you picked.
- Deploy from Cursor: connect Livops over MCPAdd Livops to ~/.cursor/mcp.json with just its address, https://livops.uz/api/mcp, and sign in through the browser. Cursor's agent can then deploy, read build logs, explain a failure and roll back, and it asks you before each tool call unless you change that.
- Deploy from Codex CLI: connect Livops over MCPAdd a [mcp_servers.livops] table with url = "https://livops.uz/api/mcp" to ~/.codex/config.toml and run codex mcp login livops. Codex can then deploy, read build logs, explain a failure and roll back, and with default_tools_approval_mode = "writes" it asks before anything that changes.
- Deploy from VS Code: Copilot agent sessions and the Livops MCP serverIn VS Code, add Livops to the workspace's .mcp.json as an http server at https://livops.uz/api/mcp, select the Copilot harness in a chat session, and complete the sign-in when prompted. Copilot can then deploy, read build logs, explain a failure and roll back, and each tool call still needs your approval.
- What Livops's AI assistant does when a build failsWhen a build fails on Livops, the previous version keeps running and one button asks the assistant why. It reads the failed build's log and your project's commands, says what went wrong and how to fix it, and may offer one fix as a button that changes nothing until you press it. It never deletes, never sees variable values, and does not edit your code.